Keynotes | K&Ts | GACs | Talks | Posters | Search
Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms
Partitioning Signal and Noise (PSN): A modality-general denoising technique for neural responses
Jacob S. Prince1, Heiko H. Schütt2, Dora Hermes3, Greta Tuckute1, Ian Charest4, Peter Brotherwood4, David G. C. Hildebrand5, Michael J. Tarr6, George A. Alvarez1, Talia Konkle1, Kendrick Kay7; 1Harvard University, 2University of Luxemburg, 3Mayo Clinic, 4Université de Montréal, 5University of Houston, 6Carnegie Mellon University, 7University of Minnesota
Presenter: Jacob S. Prince
Large-scale neural datasets typically contain only a few repeated trials per stimulus, resulting in substantial residual noise even after trial averaging. This noise limits our ability to characterize neural tuning, assess representational geometry, and evaluate computational models. We introduce Partitioning Signal and Noise (PSN), a low-rank denoising method applicable to any repeated-trial dataset. Unlike standard PCA, which retains dimensions with high total variance and therefore conflates signal with noise, PSN uses a generative model to explicitly estimate the signal and noise covariance structure of the data. This enables low-rank reconstruction that targets signal-rich dimensions while suppressing noise-dominated ones. Applied to human fMRI data from the Natural Scenes Dataset, PSN substantially improves voxel reliability, stability of ROI tuning profiles and representational geometry, and encoding model performance relative to both trial averaging and PCA denoising. Because PSN can operate on any repeated-trial data matrix, it extends naturally to intracranial EEG, scalp EEG, electrophysiology, and calcium imaging, suggesting its utility as a general-purpose tool for improving signal fidelity across neural recording modalities.
Topic Area: Methods, Tools, Theory & Neural Coding